Endothelin-1 in combination with CRB-65 enhance risk stratification in COVID-19 patients

Abstract

Background: COVID-19 continuously causes severe disease conditions and significant mortality. We evaluate whether easily accessible biomarkers can improve risk prediction of severe disease outcomes.

Methods: Our study analysed 426 COVID-19 patients collected by German CAPNETZ and PROGRESS study groups between 2020 and 2021. Troponin T high-sensitive (TnT-hs), procalcitonin (PCT), N-terminal pro brain natriuretic peptide, angiopoietin-2, copeptin, endothelin-1 (ET-1) and lipocalin-2 were measured at enrolment and related to 28d mortality/ICU admission endpoint. Logistic and relaxed LASSO regression were used to evaluate the added value of biomarkers compared to the CRB-65 score and to develop a combined risk prediction model for our endpoint.

Results: Of the 426 COVID-19 patients, 64 (15%) reached the endpoint. Among individual biomarkers, ET-1 showed the highest predictive performance (AUC = 0.76, 95% CI: 0.70-0.82). CRB-65 alone had an AUC of 0.63 (95% CI: 0.56-0.70). Our machine learning method identified CRB-65 + ET-1 to be optimal for prediction performance and model sparsity (AUC = 0.77, 95% CI: 0.71-0.83). Decision curve analysis demonstrated its greater net benefit over CRB-65 across large range of risk thresholds. The generalizability of our non-COVID CAP model (CRB-65 + TnT-hs + PCT) to COVID-19 patients was also assessed, yielding an AUC of 0.67 (95% CI: 0.60-0.74) for our primary endpoint. For 28d mortality alone as endpoint, it performed remarkably well (AUC = 0.90, 95% CI: 0.85-0.95).

Conclusion: Combining the already established clinical CRB-65 score with ET-1 significantly improves risk prediction of intensive care requirement or death within 28 days in hospitalized COVID-19 patients.

Keywords: Biomarkers; COVID-19; CRB-65; Risk prediction.

Fig. 1

Study workflow: We describe the sampling of patients, the data management and the establishment of the statistical model. The latter is based on a two-fold cross-validation procedure to estimate both, the model parameters and the hyperparameters used for training

Fig. 2

(a) Comparison of ROC curves of single biomarkers, CRB-65 alone and our selected prediction model (CRB-65 + ET-1). Respective AUCs are given in legends. (b) Decision curve analysis showing the standardized net benefit of our prediction model (black curve) and of CRB-65 alone (red curve) at different risk thresholds. Thresholds correspond to predicted probabilities of reaching the endpoint at which the clinician decides to treat the patient or to intervene in some other way